Aiki · nanobody design

Design a nanobody

Pick a target. Click your epitope. Submit. We design developability-screened VHHs that bind that face — using the integrated nanobody-design ensemble.

Anonymous ≤ $10/job · 15-30 min Magic-link ≤ $100/job · emailed in 24 hr

1 · Pick a target

Load a structure from RCSB (4-character PDB ID). UniProt + upload coming next.

try: 1LYS (lysozyme), 1UBQ (ubiquitin), 1EMA (GFP)
Load a target above to begin.

2 · Define the epitope

Click residues on the viewer to highlight them, or paste a residue list below.

Selected: 0 residues Pro tip: pick a contiguous patch of 6-15 residues — that's a typical antibody-binding face.

3 · Choose your tier

Anonymous gets a real result fast. Magic-link unlocks the full integrated ensemble.

3.5 · Pick your generators

Generators emit candidate VHH sequences against your target+epitope. Seven generators run in parallel by default; click to toggle off any you want to skip (e.g., slow ones like DiffAb ~30 min). Partner-tier generators (RFantibody, BoltzGen) are locked.

ProteinMPNN
Template redesign · target-agnostic at design · ~30s · A10G
ON
AntiFold
Antibody-specific MPNN · template redesign · target-agnostic at design · ~30s · A10G
ON
ESM-IF
General inverse folding · template redesign · target-agnostic at design · ~30s · A10G
ON
PWM-Sampler
Per-length PWM over seed CDRs · target-agnostic at design · ~1s · CPU
ON
DiffAb
CDR resampling on cAbBCII10 · currently template-only in this tier (true target-aware co-design = partner tier) · ~30 min · A10G
ON
IgGM
Joint seq + struct co-design · target-aware (uses your PDB + epitope) · ~2 min · A10G
ON
Aikium
GeNano
GDPO generative · target-aware (uses epitope peptide) · 6 rewards · ~30s · A10G
ON
RFantibody
Backbone diffusion (Nature 2025) · ~12 hr · A100
PARTNER
BoltzGen
Boltz-2 generative · ~6 hr/design · A100
PARTNER
7 of 7 free generators selected

Scorers · run on every folded candidate

Three Aikium-trained scorers re-rank the candidates that survive folding — expression (XP), solubility (Sol), and T-cell-epitope immunogenicity (HLA). Survivors are prioritised by a composite desirability index that balances all of these with fold confidence. Off → that scorer is skipped (its column shows — in the results table); generated candidates flow through fold + clustering regardless. A Boltz-2 affinity head sits behind the partner tier.

Aikium · scorer
XP
Bacterial expression prediction (ESM-C + ProtT5-XL) · CPU
ON
Aikium · scorer
Sol
Solubility prediction (ESM-2 650M, 6-output head) · CPU
ON
Aikium · scorer
HLA
T-cell-epitope immunogenicity over the CDRs (AIKI-HLA, class I+II) · CPU
ON
Boltz-2 affinity
PPI affinity head from Boltz-2 · partner tier
PARTNER
2 of 2 free scorers selected

3.6 · Choose the folding engine

Every designed candidate is folded as a target + linker + binder complex to get a per-region pLDDT and an epitope-engagement metric. ESMFold runs free on every job. Higher-accuracy folders (Boltz-2, Chai-1, AlphaFold-Multimer with MSAs) require partner-tier compute — visible here so you can see what's available; locked until we set up the engagement.

Folder choice is independent of generator choice. Switching to a partner-tier folder gives you better ranking accuracy on the same candidate pool — recommended once you have wet-lab feedback to fine-tune. Talk to Aikium →

4 · Submit

Multi-generator ensemble + multi-iteration refinement. Anonymous: 1 iteration. Magic-link: up to 3 iterations.

Each iteration reseeds the Profile PWM with top-K survivors from the prior round.
Load a target and pick an epitope to enable submit.
Production pipeline is live. Anonymous + magic-link both hit the real ensemble (ProteinMPNN + AntiFold + ESM-IF1 + Profile, in parallel) on the cAbBCII10 framework. Each candidate is folded as a target + linker + binder complex via local ESMFold (A10G) with ESMFold Atlas as fallback, scored by per-region pLDDT, and cluster-deduped at CDR3 J≥0.7. Multi-iteration mode (magic-link) seeds the Profile PWM with prior-round survivors to refine the loop ensemble.